Model reasoning and IP extraction

These are attacks that strip a model's value. By probing how the model reasons (adversarial reasoning extraction, which exposes inner chain-of-thought), an attacker gains what's needed to distil an unauthorized copy of the model, leaking its proprietary reasoning and intellectual property.